Segmentation And Classification Method Using Statistical Information For PolSAR Images

Polarimetric Synthetic aperture radar(SAR)has been widely used for target detection and recognition,extraction of surface parameters,land use and land cover classification,because it alternately transmits and receives electromagnetic wave in different polarization ways,and provide a wealth of detailed information about objects.In recent years,more and more Polarimetric SAR systems are successful launched,a large number of data sources can be afforded.It’s urgent to develop a Semi-automatic or fully automatic Pol SAR image information extraction system,which involves two critical steps,segmentation and classification.However,inherent speckle noise of Pol SAR image hinders its application on segmentation and classification.Statistical information is an inherent characteristic of speckle noise in Pol SAR images,which means the utilization of this statistic information for Pol SAR image segmentation and classification helps to reduce its interfere and improve accuracy.Hence,this study is focused on the research of Pol SAR image segmentation and classification methods using statistical information based on the theoretical analysis of Pol SAR image statistical models.In terms of Pol SAR image segmentation,this thesis systematically summarized the Pol SAR image segmentation algorithms based on statistic information,and pointed out the problems existing in the current methods,which laid a foundation for the subsequent research.To obtain appropriate initial segmentation objects,a fast and accurate simple linear iterative clustering(SLIC)algorithm was developed to generate superpixels.To utilize statistic information accurately,a regional heterogeneity index was proposed and its threshold for different Pol SAR images were determined by experiments.On this basis,a novel segmentation method,which selectively used Wishart and K distribution,was proposed.At last,the effectiveness of the proposed method was validated with both the simulated and real Pol SAR data.In terms of Pol SAR image classification,this thesis systematically summarized the existing problems in Pol SAR image classification algorithms using statistic information,and pointed out the urgency of developing object-oriented Pol SAR image classification methods.To utilize the information between the adjacent objects efficiently,the traditional pixel-based probabilistic label relaxation(PLR)procedure was developed for objects.To determine the appropriate classification scale,both the superpixels with different sizes and the segmentation results with different scales were utilized as classification elements and the influence caused by different classification scales was deeply analyzed.On this basis,a novel supervised object-oriented classification method was proposed.At last,this thesis proved the effectiveness of the proposed Pol SAR image classification method by using simulated and real Pol SAR data.The main contributions of this thesis are as follows:1)To obtain accurate segmentation areas with different heterogeneity in Pol SAR images,a novel segmentation method is proposed,which selectively uses Wishart and K statistical information based on the Fractal Network Evolution Algorithm(FNEA).Specifically,the superpixels generated by the modified Simple Linear Iterative Clustering(SLIC)algorithm are used as initial segmentation objects.Then,the similarity criterion between adjacent objects is defined by Wishart and K distribution depending on the proposed regional heterogeneity index,which contributes to the accurate utilization of statistical information in the areas with different heterogeneity.Afterwards,the similarity criterion is incorporated to the polarimetric feature generated by Pauli decomposition and the shape feature to obtain the comprehensive similarity criterion.Finally,the segmentation procedure for polarimetric data is realized,which makes full use of Wishart and K statistical information.Moreover,a simulated image,and an airborne AIRSAR image,a spaceborne RADARSAT-2 image and a spaceborne Terra SAR-X image are used to verify the effectiveness of the proposed method.The experiment result shows that this method could suppress the speckle in Pol SAR images effectually;furthermore,it can accurately segment different heterogeneity areas on the whole and get more precise boundary in the local details compared with other algorithms.2)To avoid the low accuracy and discontinuity of Pol SAR image classification results.a novel supervised superpixel-based classification method is proposed in this study to suppress the influence of speckle noise on Pol SAR images and obtain accurate and consistent classification results.This classification method combines statistical information with spatial context information based on the stochastic expectation maximization(SEM)algorithm.First,a modified simple linear iterative clustering(SLIC)algorithm is utilized to generate superpixels as classification elements.Second,class posterior probabilities of superpixels are calculated by K distribution in iterations of SEM.Then,a neighborhood function is defined to express the spatial relationship between adjacent superpixels quantitatively,and the class posterior probabilities are altered by this predefined neighborhood function in a probabilistic label relaxation(PLR)procedure.The final classification result is obtained by the maximum a posteriori decision rule,when the SEM algorithm is terminated.Several datasets,including a simulated image,a spaceborne RADARSAT-2 image,and an airborne AIRSAR image,are used to verify the validity and applicability of the proposed method.The experimental results indicate that the proposed method obtains more accurate and consistent results than other methods.